Geochemical surveys measure element concentrations in soil, stream sediments, and rock samples across large areas. Anomalous concentrations — unusually high copper, gold, arsenic — indicate potential ore deposits. The problem: “anomalous” is undefined. The background varies with geology, weathering, and transport. What's anomalous in one lithological unit is normal in another.
Chen et al. (arXiv:2603.13068) benchmark unsupervised anomaly detection methods for geochemical exploration. The key challenge is that mineral deposits are rare, spatially clustered, and have no labeled training examples — you don't know where the deposits are (that's the point of exploring). Supervised methods can't be used because there's nothing to supervise against.
The benchmark tests 35 methods across multiple geochemical datasets with known deposits. The finding: simple statistical methods (isolation forests, local outlier factors) outperform complex deep learning approaches on most datasets. The geochemical anomaly pattern is lower-dimensional than deep methods expect. A deposit produces a coherent multi-element signature (copper-gold-arsenic rising together) that doesn't require disentangling latent representations to detect — it's visible in the raw multivariate statistics.
The exception: in regions with strong lithological variation (multiple rock types with different background chemistry), the deep methods perform better because they can learn the nonlinear background model that the simple methods miss. The background IS the hard problem. Once you model the background correctly, the anomaly stands out in any method.
For mineral exploration, the practical implication is that the method matters less than the background model. Getting the background right — knowing what “normal” looks like at each location — is the load-bearing step. The anomaly detection on top is almost afterthought. The hidden deposit isn't hidden by the data; it's hidden by the geology that varies around it.